]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
7 | #include <linux/kernel.h> | |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
12 | #include <linux/kmemleak.h> | |
13 | #include <linux/mm.h> | |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
23 | #include <linux/sched/topology.h> | |
24 | #include <linux/sched/signal.h> | |
25 | #include <linux/delay.h> | |
26 | #include <linux/crash_dump.h> | |
27 | #include <linux/prefetch.h> | |
28 | ||
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
34 | #include "blk-mq-tag.h" | |
35 | #include "blk-stat.h" | |
36 | #include "blk-wbt.h" | |
37 | #include "blk-mq-sched.h" | |
38 | ||
39 | static DEFINE_MUTEX(all_q_mutex); | |
40 | static LIST_HEAD(all_q_list); | |
41 | ||
42 | static void blk_mq_poll_stats_start(struct request_queue *q); | |
43 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
44 | ||
45 | /* | |
46 | * Check if any of the ctx's have pending work in this hardware queue | |
47 | */ | |
48 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
49 | { | |
50 | return sbitmap_any_bit_set(&hctx->ctx_map) || | |
51 | !list_empty_careful(&hctx->dispatch) || | |
52 | blk_mq_sched_has_work(hctx); | |
53 | } | |
54 | ||
55 | /* | |
56 | * Mark this ctx as having pending work in this hardware queue | |
57 | */ | |
58 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
59 | struct blk_mq_ctx *ctx) | |
60 | { | |
61 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) | |
62 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
63 | } | |
64 | ||
65 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
66 | struct blk_mq_ctx *ctx) | |
67 | { | |
68 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); | |
69 | } | |
70 | ||
71 | void blk_freeze_queue_start(struct request_queue *q) | |
72 | { | |
73 | int freeze_depth; | |
74 | ||
75 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); | |
76 | if (freeze_depth == 1) { | |
77 | percpu_ref_kill(&q->q_usage_counter); | |
78 | blk_mq_run_hw_queues(q, false); | |
79 | } | |
80 | } | |
81 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); | |
82 | ||
83 | void blk_mq_freeze_queue_wait(struct request_queue *q) | |
84 | { | |
85 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); | |
86 | } | |
87 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); | |
88 | ||
89 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, | |
90 | unsigned long timeout) | |
91 | { | |
92 | return wait_event_timeout(q->mq_freeze_wq, | |
93 | percpu_ref_is_zero(&q->q_usage_counter), | |
94 | timeout); | |
95 | } | |
96 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
97 | ||
98 | /* | |
99 | * Guarantee no request is in use, so we can change any data structure of | |
100 | * the queue afterward. | |
101 | */ | |
102 | void blk_freeze_queue(struct request_queue *q) | |
103 | { | |
104 | /* | |
105 | * In the !blk_mq case we are only calling this to kill the | |
106 | * q_usage_counter, otherwise this increases the freeze depth | |
107 | * and waits for it to return to zero. For this reason there is | |
108 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
109 | * exported to drivers as the only user for unfreeze is blk_mq. | |
110 | */ | |
111 | blk_freeze_queue_start(q); | |
112 | blk_mq_freeze_queue_wait(q); | |
113 | } | |
114 | ||
115 | void blk_mq_freeze_queue(struct request_queue *q) | |
116 | { | |
117 | /* | |
118 | * ...just an alias to keep freeze and unfreeze actions balanced | |
119 | * in the blk_mq_* namespace | |
120 | */ | |
121 | blk_freeze_queue(q); | |
122 | } | |
123 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); | |
124 | ||
125 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
126 | { | |
127 | int freeze_depth; | |
128 | ||
129 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); | |
130 | WARN_ON_ONCE(freeze_depth < 0); | |
131 | if (!freeze_depth) { | |
132 | percpu_ref_reinit(&q->q_usage_counter); | |
133 | wake_up_all(&q->mq_freeze_wq); | |
134 | } | |
135 | } | |
136 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); | |
137 | ||
138 | /** | |
139 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
140 | * @q: request queue. | |
141 | * | |
142 | * Note: this function does not prevent that the struct request end_io() | |
143 | * callback function is invoked. Additionally, it is not prevented that | |
144 | * new queue_rq() calls occur unless the queue has been stopped first. | |
145 | */ | |
146 | void blk_mq_quiesce_queue(struct request_queue *q) | |
147 | { | |
148 | struct blk_mq_hw_ctx *hctx; | |
149 | unsigned int i; | |
150 | bool rcu = false; | |
151 | ||
152 | blk_mq_stop_hw_queues(q); | |
153 | ||
154 | queue_for_each_hw_ctx(q, hctx, i) { | |
155 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
156 | synchronize_srcu(&hctx->queue_rq_srcu); | |
157 | else | |
158 | rcu = true; | |
159 | } | |
160 | if (rcu) | |
161 | synchronize_rcu(); | |
162 | } | |
163 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
164 | ||
165 | void blk_mq_wake_waiters(struct request_queue *q) | |
166 | { | |
167 | struct blk_mq_hw_ctx *hctx; | |
168 | unsigned int i; | |
169 | ||
170 | queue_for_each_hw_ctx(q, hctx, i) | |
171 | if (blk_mq_hw_queue_mapped(hctx)) | |
172 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
173 | ||
174 | /* | |
175 | * If we are called because the queue has now been marked as | |
176 | * dying, we need to ensure that processes currently waiting on | |
177 | * the queue are notified as well. | |
178 | */ | |
179 | wake_up_all(&q->mq_freeze_wq); | |
180 | } | |
181 | ||
182 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) | |
183 | { | |
184 | return blk_mq_has_free_tags(hctx->tags); | |
185 | } | |
186 | EXPORT_SYMBOL(blk_mq_can_queue); | |
187 | ||
188 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, | |
189 | struct request *rq, unsigned int op) | |
190 | { | |
191 | INIT_LIST_HEAD(&rq->queuelist); | |
192 | /* csd/requeue_work/fifo_time is initialized before use */ | |
193 | rq->q = q; | |
194 | rq->mq_ctx = ctx; | |
195 | rq->cmd_flags = op; | |
196 | if (blk_queue_io_stat(q)) | |
197 | rq->rq_flags |= RQF_IO_STAT; | |
198 | /* do not touch atomic flags, it needs atomic ops against the timer */ | |
199 | rq->cpu = -1; | |
200 | INIT_HLIST_NODE(&rq->hash); | |
201 | RB_CLEAR_NODE(&rq->rb_node); | |
202 | rq->rq_disk = NULL; | |
203 | rq->part = NULL; | |
204 | rq->start_time = jiffies; | |
205 | #ifdef CONFIG_BLK_CGROUP | |
206 | rq->rl = NULL; | |
207 | set_start_time_ns(rq); | |
208 | rq->io_start_time_ns = 0; | |
209 | #endif | |
210 | rq->nr_phys_segments = 0; | |
211 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
212 | rq->nr_integrity_segments = 0; | |
213 | #endif | |
214 | rq->special = NULL; | |
215 | /* tag was already set */ | |
216 | rq->errors = 0; | |
217 | rq->extra_len = 0; | |
218 | ||
219 | INIT_LIST_HEAD(&rq->timeout_list); | |
220 | rq->timeout = 0; | |
221 | ||
222 | rq->end_io = NULL; | |
223 | rq->end_io_data = NULL; | |
224 | rq->next_rq = NULL; | |
225 | ||
226 | ctx->rq_dispatched[op_is_sync(op)]++; | |
227 | } | |
228 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); | |
229 | ||
230 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, | |
231 | unsigned int op) | |
232 | { | |
233 | struct request *rq; | |
234 | unsigned int tag; | |
235 | ||
236 | tag = blk_mq_get_tag(data); | |
237 | if (tag != BLK_MQ_TAG_FAIL) { | |
238 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); | |
239 | ||
240 | rq = tags->static_rqs[tag]; | |
241 | ||
242 | if (data->flags & BLK_MQ_REQ_INTERNAL) { | |
243 | rq->tag = -1; | |
244 | rq->internal_tag = tag; | |
245 | } else { | |
246 | if (blk_mq_tag_busy(data->hctx)) { | |
247 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
248 | atomic_inc(&data->hctx->nr_active); | |
249 | } | |
250 | rq->tag = tag; | |
251 | rq->internal_tag = -1; | |
252 | data->hctx->tags->rqs[rq->tag] = rq; | |
253 | } | |
254 | ||
255 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); | |
256 | return rq; | |
257 | } | |
258 | ||
259 | return NULL; | |
260 | } | |
261 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); | |
262 | ||
263 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, | |
264 | unsigned int flags) | |
265 | { | |
266 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; | |
267 | struct request *rq; | |
268 | int ret; | |
269 | ||
270 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); | |
271 | if (ret) | |
272 | return ERR_PTR(ret); | |
273 | ||
274 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); | |
275 | ||
276 | blk_mq_put_ctx(alloc_data.ctx); | |
277 | blk_queue_exit(q); | |
278 | ||
279 | if (!rq) | |
280 | return ERR_PTR(-EWOULDBLOCK); | |
281 | ||
282 | rq->__data_len = 0; | |
283 | rq->__sector = (sector_t) -1; | |
284 | rq->bio = rq->biotail = NULL; | |
285 | return rq; | |
286 | } | |
287 | EXPORT_SYMBOL(blk_mq_alloc_request); | |
288 | ||
289 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, | |
290 | unsigned int flags, unsigned int hctx_idx) | |
291 | { | |
292 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; | |
293 | struct request *rq; | |
294 | unsigned int cpu; | |
295 | int ret; | |
296 | ||
297 | /* | |
298 | * If the tag allocator sleeps we could get an allocation for a | |
299 | * different hardware context. No need to complicate the low level | |
300 | * allocator for this for the rare use case of a command tied to | |
301 | * a specific queue. | |
302 | */ | |
303 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
304 | return ERR_PTR(-EINVAL); | |
305 | ||
306 | if (hctx_idx >= q->nr_hw_queues) | |
307 | return ERR_PTR(-EIO); | |
308 | ||
309 | ret = blk_queue_enter(q, true); | |
310 | if (ret) | |
311 | return ERR_PTR(ret); | |
312 | ||
313 | /* | |
314 | * Check if the hardware context is actually mapped to anything. | |
315 | * If not tell the caller that it should skip this queue. | |
316 | */ | |
317 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; | |
318 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
319 | blk_queue_exit(q); | |
320 | return ERR_PTR(-EXDEV); | |
321 | } | |
322 | cpu = cpumask_first(alloc_data.hctx->cpumask); | |
323 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
324 | ||
325 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); | |
326 | ||
327 | blk_queue_exit(q); | |
328 | ||
329 | if (!rq) | |
330 | return ERR_PTR(-EWOULDBLOCK); | |
331 | ||
332 | return rq; | |
333 | } | |
334 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
335 | ||
336 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, | |
337 | struct request *rq) | |
338 | { | |
339 | const int sched_tag = rq->internal_tag; | |
340 | struct request_queue *q = rq->q; | |
341 | ||
342 | if (rq->rq_flags & RQF_MQ_INFLIGHT) | |
343 | atomic_dec(&hctx->nr_active); | |
344 | ||
345 | wbt_done(q->rq_wb, &rq->issue_stat); | |
346 | rq->rq_flags = 0; | |
347 | ||
348 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
349 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
350 | if (rq->tag != -1) | |
351 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
352 | if (sched_tag != -1) | |
353 | blk_mq_sched_completed_request(hctx, rq); | |
354 | blk_mq_sched_restart(hctx); | |
355 | blk_queue_exit(q); | |
356 | } | |
357 | ||
358 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, | |
359 | struct request *rq) | |
360 | { | |
361 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
362 | ||
363 | ctx->rq_completed[rq_is_sync(rq)]++; | |
364 | __blk_mq_finish_request(hctx, ctx, rq); | |
365 | } | |
366 | ||
367 | void blk_mq_finish_request(struct request *rq) | |
368 | { | |
369 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
370 | } | |
371 | EXPORT_SYMBOL_GPL(blk_mq_finish_request); | |
372 | ||
373 | void blk_mq_free_request(struct request *rq) | |
374 | { | |
375 | blk_mq_sched_put_request(rq); | |
376 | } | |
377 | EXPORT_SYMBOL_GPL(blk_mq_free_request); | |
378 | ||
379 | inline void __blk_mq_end_request(struct request *rq, int error) | |
380 | { | |
381 | blk_account_io_done(rq); | |
382 | ||
383 | if (rq->end_io) { | |
384 | wbt_done(rq->q->rq_wb, &rq->issue_stat); | |
385 | rq->end_io(rq, error); | |
386 | } else { | |
387 | if (unlikely(blk_bidi_rq(rq))) | |
388 | blk_mq_free_request(rq->next_rq); | |
389 | blk_mq_free_request(rq); | |
390 | } | |
391 | } | |
392 | EXPORT_SYMBOL(__blk_mq_end_request); | |
393 | ||
394 | void blk_mq_end_request(struct request *rq, int error) | |
395 | { | |
396 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
397 | BUG(); | |
398 | __blk_mq_end_request(rq, error); | |
399 | } | |
400 | EXPORT_SYMBOL(blk_mq_end_request); | |
401 | ||
402 | static void __blk_mq_complete_request_remote(void *data) | |
403 | { | |
404 | struct request *rq = data; | |
405 | ||
406 | rq->q->softirq_done_fn(rq); | |
407 | } | |
408 | ||
409 | static void blk_mq_ipi_complete_request(struct request *rq) | |
410 | { | |
411 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
412 | bool shared = false; | |
413 | int cpu; | |
414 | ||
415 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { | |
416 | rq->q->softirq_done_fn(rq); | |
417 | return; | |
418 | } | |
419 | ||
420 | cpu = get_cpu(); | |
421 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) | |
422 | shared = cpus_share_cache(cpu, ctx->cpu); | |
423 | ||
424 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
425 | rq->csd.func = __blk_mq_complete_request_remote; | |
426 | rq->csd.info = rq; | |
427 | rq->csd.flags = 0; | |
428 | smp_call_function_single_async(ctx->cpu, &rq->csd); | |
429 | } else { | |
430 | rq->q->softirq_done_fn(rq); | |
431 | } | |
432 | put_cpu(); | |
433 | } | |
434 | ||
435 | static void blk_mq_stat_add(struct request *rq) | |
436 | { | |
437 | if (rq->rq_flags & RQF_STATS) { | |
438 | blk_mq_poll_stats_start(rq->q); | |
439 | blk_stat_add(rq); | |
440 | } | |
441 | } | |
442 | ||
443 | static void __blk_mq_complete_request(struct request *rq) | |
444 | { | |
445 | struct request_queue *q = rq->q; | |
446 | ||
447 | blk_mq_stat_add(rq); | |
448 | ||
449 | if (!q->softirq_done_fn) | |
450 | blk_mq_end_request(rq, rq->errors); | |
451 | else | |
452 | blk_mq_ipi_complete_request(rq); | |
453 | } | |
454 | ||
455 | /** | |
456 | * blk_mq_complete_request - end I/O on a request | |
457 | * @rq: the request being processed | |
458 | * | |
459 | * Description: | |
460 | * Ends all I/O on a request. It does not handle partial completions. | |
461 | * The actual completion happens out-of-order, through a IPI handler. | |
462 | **/ | |
463 | void blk_mq_complete_request(struct request *rq, int error) | |
464 | { | |
465 | struct request_queue *q = rq->q; | |
466 | ||
467 | if (unlikely(blk_should_fake_timeout(q))) | |
468 | return; | |
469 | if (!blk_mark_rq_complete(rq)) { | |
470 | rq->errors = error; | |
471 | __blk_mq_complete_request(rq); | |
472 | } | |
473 | } | |
474 | EXPORT_SYMBOL(blk_mq_complete_request); | |
475 | ||
476 | int blk_mq_request_started(struct request *rq) | |
477 | { | |
478 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
479 | } | |
480 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
481 | ||
482 | void blk_mq_start_request(struct request *rq) | |
483 | { | |
484 | struct request_queue *q = rq->q; | |
485 | ||
486 | blk_mq_sched_started_request(rq); | |
487 | ||
488 | trace_block_rq_issue(q, rq); | |
489 | ||
490 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { | |
491 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); | |
492 | rq->rq_flags |= RQF_STATS; | |
493 | wbt_issue(q->rq_wb, &rq->issue_stat); | |
494 | } | |
495 | ||
496 | blk_add_timer(rq); | |
497 | ||
498 | /* | |
499 | * Ensure that ->deadline is visible before set the started | |
500 | * flag and clear the completed flag. | |
501 | */ | |
502 | smp_mb__before_atomic(); | |
503 | ||
504 | /* | |
505 | * Mark us as started and clear complete. Complete might have been | |
506 | * set if requeue raced with timeout, which then marked it as | |
507 | * complete. So be sure to clear complete again when we start | |
508 | * the request, otherwise we'll ignore the completion event. | |
509 | */ | |
510 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
511 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
512 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
513 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
514 | ||
515 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
516 | /* | |
517 | * Make sure space for the drain appears. We know we can do | |
518 | * this because max_hw_segments has been adjusted to be one | |
519 | * fewer than the device can handle. | |
520 | */ | |
521 | rq->nr_phys_segments++; | |
522 | } | |
523 | } | |
524 | EXPORT_SYMBOL(blk_mq_start_request); | |
525 | ||
526 | /* | |
527 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
528 | * flag isn't set yet, so there may be race with timeout handler, | |
529 | * but given rq->deadline is just set in .queue_rq() under | |
530 | * this situation, the race won't be possible in reality because | |
531 | * rq->timeout should be set as big enough to cover the window | |
532 | * between blk_mq_start_request() called from .queue_rq() and | |
533 | * clearing REQ_ATOM_STARTED here. | |
534 | */ | |
535 | static void __blk_mq_requeue_request(struct request *rq) | |
536 | { | |
537 | struct request_queue *q = rq->q; | |
538 | ||
539 | trace_block_rq_requeue(q, rq); | |
540 | wbt_requeue(q->rq_wb, &rq->issue_stat); | |
541 | blk_mq_sched_requeue_request(rq); | |
542 | ||
543 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { | |
544 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
545 | rq->nr_phys_segments--; | |
546 | } | |
547 | } | |
548 | ||
549 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) | |
550 | { | |
551 | __blk_mq_requeue_request(rq); | |
552 | ||
553 | BUG_ON(blk_queued_rq(rq)); | |
554 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); | |
555 | } | |
556 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
557 | ||
558 | static void blk_mq_requeue_work(struct work_struct *work) | |
559 | { | |
560 | struct request_queue *q = | |
561 | container_of(work, struct request_queue, requeue_work.work); | |
562 | LIST_HEAD(rq_list); | |
563 | struct request *rq, *next; | |
564 | unsigned long flags; | |
565 | ||
566 | spin_lock_irqsave(&q->requeue_lock, flags); | |
567 | list_splice_init(&q->requeue_list, &rq_list); | |
568 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
569 | ||
570 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
571 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) | |
572 | continue; | |
573 | ||
574 | rq->rq_flags &= ~RQF_SOFTBARRIER; | |
575 | list_del_init(&rq->queuelist); | |
576 | blk_mq_sched_insert_request(rq, true, false, false, true); | |
577 | } | |
578 | ||
579 | while (!list_empty(&rq_list)) { | |
580 | rq = list_entry(rq_list.next, struct request, queuelist); | |
581 | list_del_init(&rq->queuelist); | |
582 | blk_mq_sched_insert_request(rq, false, false, false, true); | |
583 | } | |
584 | ||
585 | blk_mq_run_hw_queues(q, false); | |
586 | } | |
587 | ||
588 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, | |
589 | bool kick_requeue_list) | |
590 | { | |
591 | struct request_queue *q = rq->q; | |
592 | unsigned long flags; | |
593 | ||
594 | /* | |
595 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
596 | * request head insertation from the workqueue. | |
597 | */ | |
598 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); | |
599 | ||
600 | spin_lock_irqsave(&q->requeue_lock, flags); | |
601 | if (at_head) { | |
602 | rq->rq_flags |= RQF_SOFTBARRIER; | |
603 | list_add(&rq->queuelist, &q->requeue_list); | |
604 | } else { | |
605 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
606 | } | |
607 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
608 | ||
609 | if (kick_requeue_list) | |
610 | blk_mq_kick_requeue_list(q); | |
611 | } | |
612 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
613 | ||
614 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
615 | { | |
616 | kblockd_schedule_delayed_work(&q->requeue_work, 0); | |
617 | } | |
618 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
619 | ||
620 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, | |
621 | unsigned long msecs) | |
622 | { | |
623 | kblockd_schedule_delayed_work(&q->requeue_work, | |
624 | msecs_to_jiffies(msecs)); | |
625 | } | |
626 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
627 | ||
628 | void blk_mq_abort_requeue_list(struct request_queue *q) | |
629 | { | |
630 | unsigned long flags; | |
631 | LIST_HEAD(rq_list); | |
632 | ||
633 | spin_lock_irqsave(&q->requeue_lock, flags); | |
634 | list_splice_init(&q->requeue_list, &rq_list); | |
635 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
636 | ||
637 | while (!list_empty(&rq_list)) { | |
638 | struct request *rq; | |
639 | ||
640 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
641 | list_del_init(&rq->queuelist); | |
642 | rq->errors = -EIO; | |
643 | blk_mq_end_request(rq, rq->errors); | |
644 | } | |
645 | } | |
646 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
647 | ||
648 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
649 | { | |
650 | if (tag < tags->nr_tags) { | |
651 | prefetch(tags->rqs[tag]); | |
652 | return tags->rqs[tag]; | |
653 | } | |
654 | ||
655 | return NULL; | |
656 | } | |
657 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
658 | ||
659 | struct blk_mq_timeout_data { | |
660 | unsigned long next; | |
661 | unsigned int next_set; | |
662 | }; | |
663 | ||
664 | void blk_mq_rq_timed_out(struct request *req, bool reserved) | |
665 | { | |
666 | const struct blk_mq_ops *ops = req->q->mq_ops; | |
667 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
668 | ||
669 | /* | |
670 | * We know that complete is set at this point. If STARTED isn't set | |
671 | * anymore, then the request isn't active and the "timeout" should | |
672 | * just be ignored. This can happen due to the bitflag ordering. | |
673 | * Timeout first checks if STARTED is set, and if it is, assumes | |
674 | * the request is active. But if we race with completion, then | |
675 | * both flags will get cleared. So check here again, and ignore | |
676 | * a timeout event with a request that isn't active. | |
677 | */ | |
678 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) | |
679 | return; | |
680 | ||
681 | if (ops->timeout) | |
682 | ret = ops->timeout(req, reserved); | |
683 | ||
684 | switch (ret) { | |
685 | case BLK_EH_HANDLED: | |
686 | __blk_mq_complete_request(req); | |
687 | break; | |
688 | case BLK_EH_RESET_TIMER: | |
689 | blk_add_timer(req); | |
690 | blk_clear_rq_complete(req); | |
691 | break; | |
692 | case BLK_EH_NOT_HANDLED: | |
693 | break; | |
694 | default: | |
695 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
696 | break; | |
697 | } | |
698 | } | |
699 | ||
700 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, | |
701 | struct request *rq, void *priv, bool reserved) | |
702 | { | |
703 | struct blk_mq_timeout_data *data = priv; | |
704 | ||
705 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) | |
706 | return; | |
707 | ||
708 | /* | |
709 | * The rq being checked may have been freed and reallocated | |
710 | * out already here, we avoid this race by checking rq->deadline | |
711 | * and REQ_ATOM_COMPLETE flag together: | |
712 | * | |
713 | * - if rq->deadline is observed as new value because of | |
714 | * reusing, the rq won't be timed out because of timing. | |
715 | * - if rq->deadline is observed as previous value, | |
716 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
717 | * because we put a barrier between setting rq->deadline | |
718 | * and clearing the flag in blk_mq_start_request(), so | |
719 | * this rq won't be timed out too. | |
720 | */ | |
721 | if (time_after_eq(jiffies, rq->deadline)) { | |
722 | if (!blk_mark_rq_complete(rq)) | |
723 | blk_mq_rq_timed_out(rq, reserved); | |
724 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { | |
725 | data->next = rq->deadline; | |
726 | data->next_set = 1; | |
727 | } | |
728 | } | |
729 | ||
730 | static void blk_mq_timeout_work(struct work_struct *work) | |
731 | { | |
732 | struct request_queue *q = | |
733 | container_of(work, struct request_queue, timeout_work); | |
734 | struct blk_mq_timeout_data data = { | |
735 | .next = 0, | |
736 | .next_set = 0, | |
737 | }; | |
738 | int i; | |
739 | ||
740 | /* A deadlock might occur if a request is stuck requiring a | |
741 | * timeout at the same time a queue freeze is waiting | |
742 | * completion, since the timeout code would not be able to | |
743 | * acquire the queue reference here. | |
744 | * | |
745 | * That's why we don't use blk_queue_enter here; instead, we use | |
746 | * percpu_ref_tryget directly, because we need to be able to | |
747 | * obtain a reference even in the short window between the queue | |
748 | * starting to freeze, by dropping the first reference in | |
749 | * blk_freeze_queue_start, and the moment the last request is | |
750 | * consumed, marked by the instant q_usage_counter reaches | |
751 | * zero. | |
752 | */ | |
753 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
754 | return; | |
755 | ||
756 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); | |
757 | ||
758 | if (data.next_set) { | |
759 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
760 | mod_timer(&q->timeout, data.next); | |
761 | } else { | |
762 | struct blk_mq_hw_ctx *hctx; | |
763 | ||
764 | queue_for_each_hw_ctx(q, hctx, i) { | |
765 | /* the hctx may be unmapped, so check it here */ | |
766 | if (blk_mq_hw_queue_mapped(hctx)) | |
767 | blk_mq_tag_idle(hctx); | |
768 | } | |
769 | } | |
770 | blk_queue_exit(q); | |
771 | } | |
772 | ||
773 | /* | |
774 | * Reverse check our software queue for entries that we could potentially | |
775 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
776 | * too much time checking for merges. | |
777 | */ | |
778 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
779 | struct blk_mq_ctx *ctx, struct bio *bio) | |
780 | { | |
781 | struct request *rq; | |
782 | int checked = 8; | |
783 | ||
784 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
785 | bool merged = false; | |
786 | ||
787 | if (!checked--) | |
788 | break; | |
789 | ||
790 | if (!blk_rq_merge_ok(rq, bio)) | |
791 | continue; | |
792 | ||
793 | switch (blk_try_merge(rq, bio)) { | |
794 | case ELEVATOR_BACK_MERGE: | |
795 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
796 | merged = bio_attempt_back_merge(q, rq, bio); | |
797 | break; | |
798 | case ELEVATOR_FRONT_MERGE: | |
799 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
800 | merged = bio_attempt_front_merge(q, rq, bio); | |
801 | break; | |
802 | case ELEVATOR_DISCARD_MERGE: | |
803 | merged = bio_attempt_discard_merge(q, rq, bio); | |
804 | break; | |
805 | default: | |
806 | continue; | |
807 | } | |
808 | ||
809 | if (merged) | |
810 | ctx->rq_merged++; | |
811 | return merged; | |
812 | } | |
813 | ||
814 | return false; | |
815 | } | |
816 | ||
817 | struct flush_busy_ctx_data { | |
818 | struct blk_mq_hw_ctx *hctx; | |
819 | struct list_head *list; | |
820 | }; | |
821 | ||
822 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
823 | { | |
824 | struct flush_busy_ctx_data *flush_data = data; | |
825 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
826 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
827 | ||
828 | sbitmap_clear_bit(sb, bitnr); | |
829 | spin_lock(&ctx->lock); | |
830 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
831 | spin_unlock(&ctx->lock); | |
832 | return true; | |
833 | } | |
834 | ||
835 | /* | |
836 | * Process software queues that have been marked busy, splicing them | |
837 | * to the for-dispatch | |
838 | */ | |
839 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
840 | { | |
841 | struct flush_busy_ctx_data data = { | |
842 | .hctx = hctx, | |
843 | .list = list, | |
844 | }; | |
845 | ||
846 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); | |
847 | } | |
848 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); | |
849 | ||
850 | static inline unsigned int queued_to_index(unsigned int queued) | |
851 | { | |
852 | if (!queued) | |
853 | return 0; | |
854 | ||
855 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); | |
856 | } | |
857 | ||
858 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, | |
859 | bool wait) | |
860 | { | |
861 | struct blk_mq_alloc_data data = { | |
862 | .q = rq->q, | |
863 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), | |
864 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
865 | }; | |
866 | ||
867 | if (rq->tag != -1) | |
868 | goto done; | |
869 | ||
870 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) | |
871 | data.flags |= BLK_MQ_REQ_RESERVED; | |
872 | ||
873 | rq->tag = blk_mq_get_tag(&data); | |
874 | if (rq->tag >= 0) { | |
875 | if (blk_mq_tag_busy(data.hctx)) { | |
876 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
877 | atomic_inc(&data.hctx->nr_active); | |
878 | } | |
879 | data.hctx->tags->rqs[rq->tag] = rq; | |
880 | } | |
881 | ||
882 | done: | |
883 | if (hctx) | |
884 | *hctx = data.hctx; | |
885 | return rq->tag != -1; | |
886 | } | |
887 | ||
888 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, | |
889 | struct request *rq) | |
890 | { | |
891 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); | |
892 | rq->tag = -1; | |
893 | ||
894 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
895 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
896 | atomic_dec(&hctx->nr_active); | |
897 | } | |
898 | } | |
899 | ||
900 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, | |
901 | struct request *rq) | |
902 | { | |
903 | if (rq->tag == -1 || rq->internal_tag == -1) | |
904 | return; | |
905 | ||
906 | __blk_mq_put_driver_tag(hctx, rq); | |
907 | } | |
908 | ||
909 | static void blk_mq_put_driver_tag(struct request *rq) | |
910 | { | |
911 | struct blk_mq_hw_ctx *hctx; | |
912 | ||
913 | if (rq->tag == -1 || rq->internal_tag == -1) | |
914 | return; | |
915 | ||
916 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
917 | __blk_mq_put_driver_tag(hctx, rq); | |
918 | } | |
919 | ||
920 | /* | |
921 | * If we fail getting a driver tag because all the driver tags are already | |
922 | * assigned and on the dispatch list, BUT the first entry does not have a | |
923 | * tag, then we could deadlock. For that case, move entries with assigned | |
924 | * driver tags to the front, leaving the set of tagged requests in the | |
925 | * same order, and the untagged set in the same order. | |
926 | */ | |
927 | static bool reorder_tags_to_front(struct list_head *list) | |
928 | { | |
929 | struct request *rq, *tmp, *first = NULL; | |
930 | ||
931 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
932 | if (rq == first) | |
933 | break; | |
934 | if (rq->tag != -1) { | |
935 | list_move(&rq->queuelist, list); | |
936 | if (!first) | |
937 | first = rq; | |
938 | } | |
939 | } | |
940 | ||
941 | return first != NULL; | |
942 | } | |
943 | ||
944 | static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags, | |
945 | void *key) | |
946 | { | |
947 | struct blk_mq_hw_ctx *hctx; | |
948 | ||
949 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
950 | ||
951 | list_del(&wait->task_list); | |
952 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); | |
953 | blk_mq_run_hw_queue(hctx, true); | |
954 | return 1; | |
955 | } | |
956 | ||
957 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
958 | { | |
959 | struct sbq_wait_state *ws; | |
960 | ||
961 | /* | |
962 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
963 | * The thread which wins the race to grab this bit adds the hardware | |
964 | * queue to the wait queue. | |
965 | */ | |
966 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
967 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
968 | return false; | |
969 | ||
970 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
971 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
972 | ||
973 | /* | |
974 | * As soon as this returns, it's no longer safe to fiddle with | |
975 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
976 | * and unlock the bit. | |
977 | */ | |
978 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
979 | return true; | |
980 | } | |
981 | ||
982 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list) | |
983 | { | |
984 | struct blk_mq_hw_ctx *hctx; | |
985 | struct request *rq; | |
986 | int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK; | |
987 | ||
988 | if (list_empty(list)) | |
989 | return false; | |
990 | ||
991 | /* | |
992 | * Now process all the entries, sending them to the driver. | |
993 | */ | |
994 | errors = queued = 0; | |
995 | do { | |
996 | struct blk_mq_queue_data bd; | |
997 | ||
998 | rq = list_first_entry(list, struct request, queuelist); | |
999 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { | |
1000 | if (!queued && reorder_tags_to_front(list)) | |
1001 | continue; | |
1002 | ||
1003 | /* | |
1004 | * The initial allocation attempt failed, so we need to | |
1005 | * rerun the hardware queue when a tag is freed. | |
1006 | */ | |
1007 | if (!blk_mq_dispatch_wait_add(hctx)) | |
1008 | break; | |
1009 | ||
1010 | /* | |
1011 | * It's possible that a tag was freed in the window | |
1012 | * between the allocation failure and adding the | |
1013 | * hardware queue to the wait queue. | |
1014 | */ | |
1015 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
1016 | break; | |
1017 | } | |
1018 | ||
1019 | list_del_init(&rq->queuelist); | |
1020 | ||
1021 | bd.rq = rq; | |
1022 | ||
1023 | /* | |
1024 | * Flag last if we have no more requests, or if we have more | |
1025 | * but can't assign a driver tag to it. | |
1026 | */ | |
1027 | if (list_empty(list)) | |
1028 | bd.last = true; | |
1029 | else { | |
1030 | struct request *nxt; | |
1031 | ||
1032 | nxt = list_first_entry(list, struct request, queuelist); | |
1033 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1034 | } | |
1035 | ||
1036 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1037 | switch (ret) { | |
1038 | case BLK_MQ_RQ_QUEUE_OK: | |
1039 | queued++; | |
1040 | break; | |
1041 | case BLK_MQ_RQ_QUEUE_BUSY: | |
1042 | blk_mq_put_driver_tag_hctx(hctx, rq); | |
1043 | list_add(&rq->queuelist, list); | |
1044 | __blk_mq_requeue_request(rq); | |
1045 | break; | |
1046 | default: | |
1047 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
1048 | case BLK_MQ_RQ_QUEUE_ERROR: | |
1049 | errors++; | |
1050 | rq->errors = -EIO; | |
1051 | blk_mq_end_request(rq, rq->errors); | |
1052 | break; | |
1053 | } | |
1054 | ||
1055 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
1056 | break; | |
1057 | } while (!list_empty(list)); | |
1058 | ||
1059 | hctx->dispatched[queued_to_index(queued)]++; | |
1060 | ||
1061 | /* | |
1062 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1063 | * that is where we will continue on next queue run. | |
1064 | */ | |
1065 | if (!list_empty(list)) { | |
1066 | /* | |
1067 | * If an I/O scheduler has been configured and we got a driver | |
1068 | * tag for the next request already, free it again. | |
1069 | */ | |
1070 | rq = list_first_entry(list, struct request, queuelist); | |
1071 | blk_mq_put_driver_tag(rq); | |
1072 | ||
1073 | spin_lock(&hctx->lock); | |
1074 | list_splice_init(list, &hctx->dispatch); | |
1075 | spin_unlock(&hctx->lock); | |
1076 | ||
1077 | /* | |
1078 | * If SCHED_RESTART was set by the caller of this function and | |
1079 | * it is no longer set that means that it was cleared by another | |
1080 | * thread and hence that a queue rerun is needed. | |
1081 | * | |
1082 | * If TAG_WAITING is set that means that an I/O scheduler has | |
1083 | * been configured and another thread is waiting for a driver | |
1084 | * tag. To guarantee fairness, do not rerun this hardware queue | |
1085 | * but let the other thread grab the driver tag. | |
1086 | * | |
1087 | * If no I/O scheduler has been configured it is possible that | |
1088 | * the hardware queue got stopped and restarted before requests | |
1089 | * were pushed back onto the dispatch list. Rerun the queue to | |
1090 | * avoid starvation. Notes: | |
1091 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1092 | * been stopped before rerunning a queue. | |
1093 | * - Some but not all block drivers stop a queue before | |
1094 | * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq | |
1095 | * and dm-rq. | |
1096 | */ | |
1097 | if (!blk_mq_sched_needs_restart(hctx) && | |
1098 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
1099 | blk_mq_run_hw_queue(hctx, true); | |
1100 | } | |
1101 | ||
1102 | return (queued + errors) != 0; | |
1103 | } | |
1104 | ||
1105 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1106 | { | |
1107 | int srcu_idx; | |
1108 | ||
1109 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1110 | cpu_online(hctx->next_cpu)); | |
1111 | ||
1112 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1113 | rcu_read_lock(); | |
1114 | blk_mq_sched_dispatch_requests(hctx); | |
1115 | rcu_read_unlock(); | |
1116 | } else { | |
1117 | might_sleep(); | |
1118 | ||
1119 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
1120 | blk_mq_sched_dispatch_requests(hctx); | |
1121 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
1122 | } | |
1123 | } | |
1124 | ||
1125 | /* | |
1126 | * It'd be great if the workqueue API had a way to pass | |
1127 | * in a mask and had some smarts for more clever placement. | |
1128 | * For now we just round-robin here, switching for every | |
1129 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1130 | */ | |
1131 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1132 | { | |
1133 | if (hctx->queue->nr_hw_queues == 1) | |
1134 | return WORK_CPU_UNBOUND; | |
1135 | ||
1136 | if (--hctx->next_cpu_batch <= 0) { | |
1137 | int next_cpu; | |
1138 | ||
1139 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1140 | if (next_cpu >= nr_cpu_ids) | |
1141 | next_cpu = cpumask_first(hctx->cpumask); | |
1142 | ||
1143 | hctx->next_cpu = next_cpu; | |
1144 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1145 | } | |
1146 | ||
1147 | return hctx->next_cpu; | |
1148 | } | |
1149 | ||
1150 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, | |
1151 | unsigned long msecs) | |
1152 | { | |
1153 | if (unlikely(blk_mq_hctx_stopped(hctx) || | |
1154 | !blk_mq_hw_queue_mapped(hctx))) | |
1155 | return; | |
1156 | ||
1157 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1158 | int cpu = get_cpu(); | |
1159 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
1160 | __blk_mq_run_hw_queue(hctx); | |
1161 | put_cpu(); | |
1162 | return; | |
1163 | } | |
1164 | ||
1165 | put_cpu(); | |
1166 | } | |
1167 | ||
1168 | if (msecs == 0) | |
1169 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), | |
1170 | &hctx->run_work); | |
1171 | else | |
1172 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1173 | &hctx->delayed_run_work, | |
1174 | msecs_to_jiffies(msecs)); | |
1175 | } | |
1176 | ||
1177 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1178 | { | |
1179 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1180 | } | |
1181 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1182 | ||
1183 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1184 | { | |
1185 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
1186 | } | |
1187 | EXPORT_SYMBOL(blk_mq_run_hw_queue); | |
1188 | ||
1189 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) | |
1190 | { | |
1191 | struct blk_mq_hw_ctx *hctx; | |
1192 | int i; | |
1193 | ||
1194 | queue_for_each_hw_ctx(q, hctx, i) { | |
1195 | if (!blk_mq_hctx_has_pending(hctx) || | |
1196 | blk_mq_hctx_stopped(hctx)) | |
1197 | continue; | |
1198 | ||
1199 | blk_mq_run_hw_queue(hctx, async); | |
1200 | } | |
1201 | } | |
1202 | EXPORT_SYMBOL(blk_mq_run_hw_queues); | |
1203 | ||
1204 | /** | |
1205 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1206 | * @q: request queue. | |
1207 | * | |
1208 | * The caller is responsible for serializing this function against | |
1209 | * blk_mq_{start,stop}_hw_queue(). | |
1210 | */ | |
1211 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1212 | { | |
1213 | struct blk_mq_hw_ctx *hctx; | |
1214 | int i; | |
1215 | ||
1216 | queue_for_each_hw_ctx(q, hctx, i) | |
1217 | if (blk_mq_hctx_stopped(hctx)) | |
1218 | return true; | |
1219 | ||
1220 | return false; | |
1221 | } | |
1222 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1223 | ||
1224 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1225 | { | |
1226 | cancel_work(&hctx->run_work); | |
1227 | cancel_delayed_work(&hctx->delay_work); | |
1228 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1229 | } | |
1230 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1231 | ||
1232 | void blk_mq_stop_hw_queues(struct request_queue *q) | |
1233 | { | |
1234 | struct blk_mq_hw_ctx *hctx; | |
1235 | int i; | |
1236 | ||
1237 | queue_for_each_hw_ctx(q, hctx, i) | |
1238 | blk_mq_stop_hw_queue(hctx); | |
1239 | } | |
1240 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1241 | ||
1242 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1243 | { | |
1244 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1245 | ||
1246 | blk_mq_run_hw_queue(hctx, false); | |
1247 | } | |
1248 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1249 | ||
1250 | void blk_mq_start_hw_queues(struct request_queue *q) | |
1251 | { | |
1252 | struct blk_mq_hw_ctx *hctx; | |
1253 | int i; | |
1254 | ||
1255 | queue_for_each_hw_ctx(q, hctx, i) | |
1256 | blk_mq_start_hw_queue(hctx); | |
1257 | } | |
1258 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1259 | ||
1260 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1261 | { | |
1262 | if (!blk_mq_hctx_stopped(hctx)) | |
1263 | return; | |
1264 | ||
1265 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1266 | blk_mq_run_hw_queue(hctx, async); | |
1267 | } | |
1268 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1269 | ||
1270 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) | |
1271 | { | |
1272 | struct blk_mq_hw_ctx *hctx; | |
1273 | int i; | |
1274 | ||
1275 | queue_for_each_hw_ctx(q, hctx, i) | |
1276 | blk_mq_start_stopped_hw_queue(hctx, async); | |
1277 | } | |
1278 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1279 | ||
1280 | static void blk_mq_run_work_fn(struct work_struct *work) | |
1281 | { | |
1282 | struct blk_mq_hw_ctx *hctx; | |
1283 | ||
1284 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); | |
1285 | ||
1286 | __blk_mq_run_hw_queue(hctx); | |
1287 | } | |
1288 | ||
1289 | static void blk_mq_delayed_run_work_fn(struct work_struct *work) | |
1290 | { | |
1291 | struct blk_mq_hw_ctx *hctx; | |
1292 | ||
1293 | hctx = container_of(work, struct blk_mq_hw_ctx, delayed_run_work.work); | |
1294 | ||
1295 | __blk_mq_run_hw_queue(hctx); | |
1296 | } | |
1297 | ||
1298 | static void blk_mq_delay_work_fn(struct work_struct *work) | |
1299 | { | |
1300 | struct blk_mq_hw_ctx *hctx; | |
1301 | ||
1302 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1303 | ||
1304 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1305 | __blk_mq_run_hw_queue(hctx); | |
1306 | } | |
1307 | ||
1308 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1309 | { | |
1310 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) | |
1311 | return; | |
1312 | ||
1313 | blk_mq_stop_hw_queue(hctx); | |
1314 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1315 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
1316 | } | |
1317 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1318 | ||
1319 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, | |
1320 | struct request *rq, | |
1321 | bool at_head) | |
1322 | { | |
1323 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1324 | ||
1325 | trace_block_rq_insert(hctx->queue, rq); | |
1326 | ||
1327 | if (at_head) | |
1328 | list_add(&rq->queuelist, &ctx->rq_list); | |
1329 | else | |
1330 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
1331 | } | |
1332 | ||
1333 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, | |
1334 | bool at_head) | |
1335 | { | |
1336 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1337 | ||
1338 | __blk_mq_insert_req_list(hctx, rq, at_head); | |
1339 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1340 | } | |
1341 | ||
1342 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, | |
1343 | struct list_head *list) | |
1344 | ||
1345 | { | |
1346 | /* | |
1347 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1348 | * offline now | |
1349 | */ | |
1350 | spin_lock(&ctx->lock); | |
1351 | while (!list_empty(list)) { | |
1352 | struct request *rq; | |
1353 | ||
1354 | rq = list_first_entry(list, struct request, queuelist); | |
1355 | BUG_ON(rq->mq_ctx != ctx); | |
1356 | list_del_init(&rq->queuelist); | |
1357 | __blk_mq_insert_req_list(hctx, rq, false); | |
1358 | } | |
1359 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1360 | spin_unlock(&ctx->lock); | |
1361 | } | |
1362 | ||
1363 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1364 | { | |
1365 | struct request *rqa = container_of(a, struct request, queuelist); | |
1366 | struct request *rqb = container_of(b, struct request, queuelist); | |
1367 | ||
1368 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1369 | (rqa->mq_ctx == rqb->mq_ctx && | |
1370 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1371 | } | |
1372 | ||
1373 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1374 | { | |
1375 | struct blk_mq_ctx *this_ctx; | |
1376 | struct request_queue *this_q; | |
1377 | struct request *rq; | |
1378 | LIST_HEAD(list); | |
1379 | LIST_HEAD(ctx_list); | |
1380 | unsigned int depth; | |
1381 | ||
1382 | list_splice_init(&plug->mq_list, &list); | |
1383 | ||
1384 | list_sort(NULL, &list, plug_ctx_cmp); | |
1385 | ||
1386 | this_q = NULL; | |
1387 | this_ctx = NULL; | |
1388 | depth = 0; | |
1389 | ||
1390 | while (!list_empty(&list)) { | |
1391 | rq = list_entry_rq(list.next); | |
1392 | list_del_init(&rq->queuelist); | |
1393 | BUG_ON(!rq->q); | |
1394 | if (rq->mq_ctx != this_ctx) { | |
1395 | if (this_ctx) { | |
1396 | trace_block_unplug(this_q, depth, from_schedule); | |
1397 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1398 | &ctx_list, | |
1399 | from_schedule); | |
1400 | } | |
1401 | ||
1402 | this_ctx = rq->mq_ctx; | |
1403 | this_q = rq->q; | |
1404 | depth = 0; | |
1405 | } | |
1406 | ||
1407 | depth++; | |
1408 | list_add_tail(&rq->queuelist, &ctx_list); | |
1409 | } | |
1410 | ||
1411 | /* | |
1412 | * If 'this_ctx' is set, we know we have entries to complete | |
1413 | * on 'ctx_list'. Do those. | |
1414 | */ | |
1415 | if (this_ctx) { | |
1416 | trace_block_unplug(this_q, depth, from_schedule); | |
1417 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1418 | from_schedule); | |
1419 | } | |
1420 | } | |
1421 | ||
1422 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1423 | { | |
1424 | init_request_from_bio(rq, bio); | |
1425 | ||
1426 | blk_account_io_start(rq, true); | |
1427 | } | |
1428 | ||
1429 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) | |
1430 | { | |
1431 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1432 | !blk_queue_nomerges(hctx->queue); | |
1433 | } | |
1434 | ||
1435 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, | |
1436 | struct blk_mq_ctx *ctx, | |
1437 | struct request *rq, struct bio *bio) | |
1438 | { | |
1439 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { | |
1440 | blk_mq_bio_to_request(rq, bio); | |
1441 | spin_lock(&ctx->lock); | |
1442 | insert_rq: | |
1443 | __blk_mq_insert_request(hctx, rq, false); | |
1444 | spin_unlock(&ctx->lock); | |
1445 | return false; | |
1446 | } else { | |
1447 | struct request_queue *q = hctx->queue; | |
1448 | ||
1449 | spin_lock(&ctx->lock); | |
1450 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1451 | blk_mq_bio_to_request(rq, bio); | |
1452 | goto insert_rq; | |
1453 | } | |
1454 | ||
1455 | spin_unlock(&ctx->lock); | |
1456 | __blk_mq_finish_request(hctx, ctx, rq); | |
1457 | return true; | |
1458 | } | |
1459 | } | |
1460 | ||
1461 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) | |
1462 | { | |
1463 | if (rq->tag != -1) | |
1464 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1465 | ||
1466 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
1467 | } | |
1468 | ||
1469 | static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie, | |
1470 | bool may_sleep) | |
1471 | { | |
1472 | struct request_queue *q = rq->q; | |
1473 | struct blk_mq_queue_data bd = { | |
1474 | .rq = rq, | |
1475 | .last = true, | |
1476 | }; | |
1477 | struct blk_mq_hw_ctx *hctx; | |
1478 | blk_qc_t new_cookie; | |
1479 | int ret; | |
1480 | ||
1481 | if (q->elevator) | |
1482 | goto insert; | |
1483 | ||
1484 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
1485 | goto insert; | |
1486 | ||
1487 | new_cookie = request_to_qc_t(hctx, rq); | |
1488 | ||
1489 | /* | |
1490 | * For OK queue, we are done. For error, kill it. Any other | |
1491 | * error (busy), just add it to our list as we previously | |
1492 | * would have done | |
1493 | */ | |
1494 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1495 | if (ret == BLK_MQ_RQ_QUEUE_OK) { | |
1496 | *cookie = new_cookie; | |
1497 | return; | |
1498 | } | |
1499 | ||
1500 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1501 | *cookie = BLK_QC_T_NONE; | |
1502 | rq->errors = -EIO; | |
1503 | blk_mq_end_request(rq, rq->errors); | |
1504 | return; | |
1505 | } | |
1506 | ||
1507 | __blk_mq_requeue_request(rq); | |
1508 | insert: | |
1509 | blk_mq_sched_insert_request(rq, false, true, false, may_sleep); | |
1510 | } | |
1511 | ||
1512 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1513 | struct request *rq, blk_qc_t *cookie) | |
1514 | { | |
1515 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1516 | rcu_read_lock(); | |
1517 | __blk_mq_try_issue_directly(rq, cookie, false); | |
1518 | rcu_read_unlock(); | |
1519 | } else { | |
1520 | unsigned int srcu_idx; | |
1521 | ||
1522 | might_sleep(); | |
1523 | ||
1524 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
1525 | __blk_mq_try_issue_directly(rq, cookie, true); | |
1526 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
1527 | } | |
1528 | } | |
1529 | ||
1530 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) | |
1531 | { | |
1532 | const int is_sync = op_is_sync(bio->bi_opf); | |
1533 | const int is_flush_fua = op_is_flush(bio->bi_opf); | |
1534 | struct blk_mq_alloc_data data = { .flags = 0 }; | |
1535 | struct request *rq; | |
1536 | unsigned int request_count = 0; | |
1537 | struct blk_plug *plug; | |
1538 | struct request *same_queue_rq = NULL; | |
1539 | blk_qc_t cookie; | |
1540 | unsigned int wb_acct; | |
1541 | ||
1542 | blk_queue_bounce(q, &bio); | |
1543 | ||
1544 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
1545 | bio_io_error(bio); | |
1546 | return BLK_QC_T_NONE; | |
1547 | } | |
1548 | ||
1549 | blk_queue_split(q, &bio, q->bio_split); | |
1550 | ||
1551 | if (!is_flush_fua && !blk_queue_nomerges(q) && | |
1552 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1553 | return BLK_QC_T_NONE; | |
1554 | ||
1555 | if (blk_mq_sched_bio_merge(q, bio)) | |
1556 | return BLK_QC_T_NONE; | |
1557 | ||
1558 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); | |
1559 | ||
1560 | trace_block_getrq(q, bio, bio->bi_opf); | |
1561 | ||
1562 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
1563 | if (unlikely(!rq)) { | |
1564 | __wbt_done(q->rq_wb, wb_acct); | |
1565 | return BLK_QC_T_NONE; | |
1566 | } | |
1567 | ||
1568 | wbt_track(&rq->issue_stat, wb_acct); | |
1569 | ||
1570 | cookie = request_to_qc_t(data.hctx, rq); | |
1571 | ||
1572 | plug = current->plug; | |
1573 | if (unlikely(is_flush_fua)) { | |
1574 | blk_mq_bio_to_request(rq, bio); | |
1575 | if (q->elevator) { | |
1576 | blk_mq_sched_insert_request(rq, false, true, true, | |
1577 | true); | |
1578 | } else { | |
1579 | blk_insert_flush(rq); | |
1580 | blk_mq_run_hw_queue(data.hctx, true); | |
1581 | } | |
1582 | } else if (plug && q->nr_hw_queues == 1) { | |
1583 | struct request *last = NULL; | |
1584 | ||
1585 | blk_mq_bio_to_request(rq, bio); | |
1586 | ||
1587 | /* | |
1588 | * @request_count may become stale because of schedule | |
1589 | * out, so check the list again. | |
1590 | */ | |
1591 | if (list_empty(&plug->mq_list)) | |
1592 | request_count = 0; | |
1593 | else if (blk_queue_nomerges(q)) | |
1594 | request_count = blk_plug_queued_count(q); | |
1595 | ||
1596 | if (!request_count) | |
1597 | trace_block_plug(q); | |
1598 | else | |
1599 | last = list_entry_rq(plug->mq_list.prev); | |
1600 | ||
1601 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && | |
1602 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1603 | blk_flush_plug_list(plug, false); | |
1604 | trace_block_plug(q); | |
1605 | } | |
1606 | ||
1607 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1608 | } else if (plug && !blk_queue_nomerges(q)) { | |
1609 | blk_mq_bio_to_request(rq, bio); | |
1610 | ||
1611 | /* | |
1612 | * We do limited plugging. If the bio can be merged, do that. | |
1613 | * Otherwise the existing request in the plug list will be | |
1614 | * issued. So the plug list will have one request at most | |
1615 | * The plug list might get flushed before this. If that happens, | |
1616 | * the plug list is empty, and same_queue_rq is invalid. | |
1617 | */ | |
1618 | if (list_empty(&plug->mq_list)) | |
1619 | same_queue_rq = NULL; | |
1620 | if (same_queue_rq) | |
1621 | list_del_init(&same_queue_rq->queuelist); | |
1622 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1623 | ||
1624 | blk_mq_put_ctx(data.ctx); | |
1625 | ||
1626 | if (same_queue_rq) | |
1627 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, | |
1628 | &cookie); | |
1629 | ||
1630 | return cookie; | |
1631 | } else if (q->nr_hw_queues > 1 && is_sync) { | |
1632 | blk_mq_put_ctx(data.ctx); | |
1633 | blk_mq_bio_to_request(rq, bio); | |
1634 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); | |
1635 | return cookie; | |
1636 | } else if (q->elevator) { | |
1637 | blk_mq_bio_to_request(rq, bio); | |
1638 | blk_mq_sched_insert_request(rq, false, true, true, true); | |
1639 | } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) | |
1640 | blk_mq_run_hw_queue(data.hctx, true); | |
1641 | ||
1642 | blk_mq_put_ctx(data.ctx); | |
1643 | return cookie; | |
1644 | } | |
1645 | ||
1646 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1647 | unsigned int hctx_idx) | |
1648 | { | |
1649 | struct page *page; | |
1650 | ||
1651 | if (tags->rqs && set->ops->exit_request) { | |
1652 | int i; | |
1653 | ||
1654 | for (i = 0; i < tags->nr_tags; i++) { | |
1655 | struct request *rq = tags->static_rqs[i]; | |
1656 | ||
1657 | if (!rq) | |
1658 | continue; | |
1659 | set->ops->exit_request(set->driver_data, rq, | |
1660 | hctx_idx, i); | |
1661 | tags->static_rqs[i] = NULL; | |
1662 | } | |
1663 | } | |
1664 | ||
1665 | while (!list_empty(&tags->page_list)) { | |
1666 | page = list_first_entry(&tags->page_list, struct page, lru); | |
1667 | list_del_init(&page->lru); | |
1668 | /* | |
1669 | * Remove kmemleak object previously allocated in | |
1670 | * blk_mq_init_rq_map(). | |
1671 | */ | |
1672 | kmemleak_free(page_address(page)); | |
1673 | __free_pages(page, page->private); | |
1674 | } | |
1675 | } | |
1676 | ||
1677 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) | |
1678 | { | |
1679 | kfree(tags->rqs); | |
1680 | tags->rqs = NULL; | |
1681 | kfree(tags->static_rqs); | |
1682 | tags->static_rqs = NULL; | |
1683 | ||
1684 | blk_mq_free_tags(tags); | |
1685 | } | |
1686 | ||
1687 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, | |
1688 | unsigned int hctx_idx, | |
1689 | unsigned int nr_tags, | |
1690 | unsigned int reserved_tags) | |
1691 | { | |
1692 | struct blk_mq_tags *tags; | |
1693 | int node; | |
1694 | ||
1695 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1696 | if (node == NUMA_NO_NODE) | |
1697 | node = set->numa_node; | |
1698 | ||
1699 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
1700 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
1701 | if (!tags) | |
1702 | return NULL; | |
1703 | ||
1704 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), | |
1705 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1706 | node); | |
1707 | if (!tags->rqs) { | |
1708 | blk_mq_free_tags(tags); | |
1709 | return NULL; | |
1710 | } | |
1711 | ||
1712 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), | |
1713 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1714 | node); | |
1715 | if (!tags->static_rqs) { | |
1716 | kfree(tags->rqs); | |
1717 | blk_mq_free_tags(tags); | |
1718 | return NULL; | |
1719 | } | |
1720 | ||
1721 | return tags; | |
1722 | } | |
1723 | ||
1724 | static size_t order_to_size(unsigned int order) | |
1725 | { | |
1726 | return (size_t)PAGE_SIZE << order; | |
1727 | } | |
1728 | ||
1729 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1730 | unsigned int hctx_idx, unsigned int depth) | |
1731 | { | |
1732 | unsigned int i, j, entries_per_page, max_order = 4; | |
1733 | size_t rq_size, left; | |
1734 | int node; | |
1735 | ||
1736 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1737 | if (node == NUMA_NO_NODE) | |
1738 | node = set->numa_node; | |
1739 | ||
1740 | INIT_LIST_HEAD(&tags->page_list); | |
1741 | ||
1742 | /* | |
1743 | * rq_size is the size of the request plus driver payload, rounded | |
1744 | * to the cacheline size | |
1745 | */ | |
1746 | rq_size = round_up(sizeof(struct request) + set->cmd_size, | |
1747 | cache_line_size()); | |
1748 | left = rq_size * depth; | |
1749 | ||
1750 | for (i = 0; i < depth; ) { | |
1751 | int this_order = max_order; | |
1752 | struct page *page; | |
1753 | int to_do; | |
1754 | void *p; | |
1755 | ||
1756 | while (this_order && left < order_to_size(this_order - 1)) | |
1757 | this_order--; | |
1758 | ||
1759 | do { | |
1760 | page = alloc_pages_node(node, | |
1761 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, | |
1762 | this_order); | |
1763 | if (page) | |
1764 | break; | |
1765 | if (!this_order--) | |
1766 | break; | |
1767 | if (order_to_size(this_order) < rq_size) | |
1768 | break; | |
1769 | } while (1); | |
1770 | ||
1771 | if (!page) | |
1772 | goto fail; | |
1773 | ||
1774 | page->private = this_order; | |
1775 | list_add_tail(&page->lru, &tags->page_list); | |
1776 | ||
1777 | p = page_address(page); | |
1778 | /* | |
1779 | * Allow kmemleak to scan these pages as they contain pointers | |
1780 | * to additional allocations like via ops->init_request(). | |
1781 | */ | |
1782 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); | |
1783 | entries_per_page = order_to_size(this_order) / rq_size; | |
1784 | to_do = min(entries_per_page, depth - i); | |
1785 | left -= to_do * rq_size; | |
1786 | for (j = 0; j < to_do; j++) { | |
1787 | struct request *rq = p; | |
1788 | ||
1789 | tags->static_rqs[i] = rq; | |
1790 | if (set->ops->init_request) { | |
1791 | if (set->ops->init_request(set->driver_data, | |
1792 | rq, hctx_idx, i, | |
1793 | node)) { | |
1794 | tags->static_rqs[i] = NULL; | |
1795 | goto fail; | |
1796 | } | |
1797 | } | |
1798 | ||
1799 | p += rq_size; | |
1800 | i++; | |
1801 | } | |
1802 | } | |
1803 | return 0; | |
1804 | ||
1805 | fail: | |
1806 | blk_mq_free_rqs(set, tags, hctx_idx); | |
1807 | return -ENOMEM; | |
1808 | } | |
1809 | ||
1810 | /* | |
1811 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1812 | * software queue to the hw queue dispatch list, and ensure that it | |
1813 | * gets run. | |
1814 | */ | |
1815 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) | |
1816 | { | |
1817 | struct blk_mq_hw_ctx *hctx; | |
1818 | struct blk_mq_ctx *ctx; | |
1819 | LIST_HEAD(tmp); | |
1820 | ||
1821 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); | |
1822 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); | |
1823 | ||
1824 | spin_lock(&ctx->lock); | |
1825 | if (!list_empty(&ctx->rq_list)) { | |
1826 | list_splice_init(&ctx->rq_list, &tmp); | |
1827 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1828 | } | |
1829 | spin_unlock(&ctx->lock); | |
1830 | ||
1831 | if (list_empty(&tmp)) | |
1832 | return 0; | |
1833 | ||
1834 | spin_lock(&hctx->lock); | |
1835 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1836 | spin_unlock(&hctx->lock); | |
1837 | ||
1838 | blk_mq_run_hw_queue(hctx, true); | |
1839 | return 0; | |
1840 | } | |
1841 | ||
1842 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) | |
1843 | { | |
1844 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, | |
1845 | &hctx->cpuhp_dead); | |
1846 | } | |
1847 | ||
1848 | /* hctx->ctxs will be freed in queue's release handler */ | |
1849 | static void blk_mq_exit_hctx(struct request_queue *q, | |
1850 | struct blk_mq_tag_set *set, | |
1851 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1852 | { | |
1853 | unsigned flush_start_tag = set->queue_depth; | |
1854 | ||
1855 | blk_mq_tag_idle(hctx); | |
1856 | ||
1857 | if (set->ops->exit_request) | |
1858 | set->ops->exit_request(set->driver_data, | |
1859 | hctx->fq->flush_rq, hctx_idx, | |
1860 | flush_start_tag + hctx_idx); | |
1861 | ||
1862 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
1863 | ||
1864 | if (set->ops->exit_hctx) | |
1865 | set->ops->exit_hctx(hctx, hctx_idx); | |
1866 | ||
1867 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
1868 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1869 | ||
1870 | blk_mq_remove_cpuhp(hctx); | |
1871 | blk_free_flush_queue(hctx->fq); | |
1872 | sbitmap_free(&hctx->ctx_map); | |
1873 | } | |
1874 | ||
1875 | static void blk_mq_exit_hw_queues(struct request_queue *q, | |
1876 | struct blk_mq_tag_set *set, int nr_queue) | |
1877 | { | |
1878 | struct blk_mq_hw_ctx *hctx; | |
1879 | unsigned int i; | |
1880 | ||
1881 | queue_for_each_hw_ctx(q, hctx, i) { | |
1882 | if (i == nr_queue) | |
1883 | break; | |
1884 | blk_mq_exit_hctx(q, set, hctx, i); | |
1885 | } | |
1886 | } | |
1887 | ||
1888 | static int blk_mq_init_hctx(struct request_queue *q, | |
1889 | struct blk_mq_tag_set *set, | |
1890 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
1891 | { | |
1892 | int node; | |
1893 | unsigned flush_start_tag = set->queue_depth; | |
1894 | ||
1895 | node = hctx->numa_node; | |
1896 | if (node == NUMA_NO_NODE) | |
1897 | node = hctx->numa_node = set->numa_node; | |
1898 | ||
1899 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
1900 | INIT_DELAYED_WORK(&hctx->delayed_run_work, blk_mq_delayed_run_work_fn); | |
1901 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
1902 | spin_lock_init(&hctx->lock); | |
1903 | INIT_LIST_HEAD(&hctx->dispatch); | |
1904 | hctx->queue = q; | |
1905 | hctx->queue_num = hctx_idx; | |
1906 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; | |
1907 | ||
1908 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
1909 | ||
1910 | hctx->tags = set->tags[hctx_idx]; | |
1911 | ||
1912 | /* | |
1913 | * Allocate space for all possible cpus to avoid allocation at | |
1914 | * runtime | |
1915 | */ | |
1916 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), | |
1917 | GFP_KERNEL, node); | |
1918 | if (!hctx->ctxs) | |
1919 | goto unregister_cpu_notifier; | |
1920 | ||
1921 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, | |
1922 | node)) | |
1923 | goto free_ctxs; | |
1924 | ||
1925 | hctx->nr_ctx = 0; | |
1926 | ||
1927 | if (set->ops->init_hctx && | |
1928 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1929 | goto free_bitmap; | |
1930 | ||
1931 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) | |
1932 | goto exit_hctx; | |
1933 | ||
1934 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); | |
1935 | if (!hctx->fq) | |
1936 | goto sched_exit_hctx; | |
1937 | ||
1938 | if (set->ops->init_request && | |
1939 | set->ops->init_request(set->driver_data, | |
1940 | hctx->fq->flush_rq, hctx_idx, | |
1941 | flush_start_tag + hctx_idx, node)) | |
1942 | goto free_fq; | |
1943 | ||
1944 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
1945 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1946 | ||
1947 | return 0; | |
1948 | ||
1949 | free_fq: | |
1950 | kfree(hctx->fq); | |
1951 | sched_exit_hctx: | |
1952 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
1953 | exit_hctx: | |
1954 | if (set->ops->exit_hctx) | |
1955 | set->ops->exit_hctx(hctx, hctx_idx); | |
1956 | free_bitmap: | |
1957 | sbitmap_free(&hctx->ctx_map); | |
1958 | free_ctxs: | |
1959 | kfree(hctx->ctxs); | |
1960 | unregister_cpu_notifier: | |
1961 | blk_mq_remove_cpuhp(hctx); | |
1962 | return -1; | |
1963 | } | |
1964 | ||
1965 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
1966 | unsigned int nr_hw_queues) | |
1967 | { | |
1968 | unsigned int i; | |
1969 | ||
1970 | for_each_possible_cpu(i) { | |
1971 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1972 | struct blk_mq_hw_ctx *hctx; | |
1973 | ||
1974 | __ctx->cpu = i; | |
1975 | spin_lock_init(&__ctx->lock); | |
1976 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1977 | __ctx->queue = q; | |
1978 | ||
1979 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
1980 | if (!cpu_online(i)) | |
1981 | continue; | |
1982 | ||
1983 | hctx = blk_mq_map_queue(q, i); | |
1984 | ||
1985 | /* | |
1986 | * Set local node, IFF we have more than one hw queue. If | |
1987 | * not, we remain on the home node of the device | |
1988 | */ | |
1989 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
1990 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
1991 | } | |
1992 | } | |
1993 | ||
1994 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) | |
1995 | { | |
1996 | int ret = 0; | |
1997 | ||
1998 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1999 | set->queue_depth, set->reserved_tags); | |
2000 | if (!set->tags[hctx_idx]) | |
2001 | return false; | |
2002 | ||
2003 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2004 | set->queue_depth); | |
2005 | if (!ret) | |
2006 | return true; | |
2007 | ||
2008 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2009 | set->tags[hctx_idx] = NULL; | |
2010 | return false; | |
2011 | } | |
2012 | ||
2013 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2014 | unsigned int hctx_idx) | |
2015 | { | |
2016 | if (set->tags[hctx_idx]) { | |
2017 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2018 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2019 | set->tags[hctx_idx] = NULL; | |
2020 | } | |
2021 | } | |
2022 | ||
2023 | static void blk_mq_map_swqueue(struct request_queue *q, | |
2024 | const struct cpumask *online_mask) | |
2025 | { | |
2026 | unsigned int i, hctx_idx; | |
2027 | struct blk_mq_hw_ctx *hctx; | |
2028 | struct blk_mq_ctx *ctx; | |
2029 | struct blk_mq_tag_set *set = q->tag_set; | |
2030 | ||
2031 | /* | |
2032 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2033 | */ | |
2034 | mutex_lock(&q->sysfs_lock); | |
2035 | ||
2036 | queue_for_each_hw_ctx(q, hctx, i) { | |
2037 | cpumask_clear(hctx->cpumask); | |
2038 | hctx->nr_ctx = 0; | |
2039 | } | |
2040 | ||
2041 | /* | |
2042 | * Map software to hardware queues | |
2043 | */ | |
2044 | for_each_possible_cpu(i) { | |
2045 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
2046 | if (!cpumask_test_cpu(i, online_mask)) | |
2047 | continue; | |
2048 | ||
2049 | hctx_idx = q->mq_map[i]; | |
2050 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2051 | if (!set->tags[hctx_idx] && | |
2052 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2053 | /* | |
2054 | * If tags initialization fail for some hctx, | |
2055 | * that hctx won't be brought online. In this | |
2056 | * case, remap the current ctx to hctx[0] which | |
2057 | * is guaranteed to always have tags allocated | |
2058 | */ | |
2059 | q->mq_map[i] = 0; | |
2060 | } | |
2061 | ||
2062 | ctx = per_cpu_ptr(q->queue_ctx, i); | |
2063 | hctx = blk_mq_map_queue(q, i); | |
2064 | ||
2065 | cpumask_set_cpu(i, hctx->cpumask); | |
2066 | ctx->index_hw = hctx->nr_ctx; | |
2067 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2068 | } | |
2069 | ||
2070 | mutex_unlock(&q->sysfs_lock); | |
2071 | ||
2072 | queue_for_each_hw_ctx(q, hctx, i) { | |
2073 | /* | |
2074 | * If no software queues are mapped to this hardware queue, | |
2075 | * disable it and free the request entries. | |
2076 | */ | |
2077 | if (!hctx->nr_ctx) { | |
2078 | /* Never unmap queue 0. We need it as a | |
2079 | * fallback in case of a new remap fails | |
2080 | * allocation | |
2081 | */ | |
2082 | if (i && set->tags[i]) | |
2083 | blk_mq_free_map_and_requests(set, i); | |
2084 | ||
2085 | hctx->tags = NULL; | |
2086 | continue; | |
2087 | } | |
2088 | ||
2089 | hctx->tags = set->tags[i]; | |
2090 | WARN_ON(!hctx->tags); | |
2091 | ||
2092 | /* | |
2093 | * Set the map size to the number of mapped software queues. | |
2094 | * This is more accurate and more efficient than looping | |
2095 | * over all possibly mapped software queues. | |
2096 | */ | |
2097 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); | |
2098 | ||
2099 | /* | |
2100 | * Initialize batch roundrobin counts | |
2101 | */ | |
2102 | hctx->next_cpu = cpumask_first(hctx->cpumask); | |
2103 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2104 | } | |
2105 | } | |
2106 | ||
2107 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) | |
2108 | { | |
2109 | struct blk_mq_hw_ctx *hctx; | |
2110 | int i; | |
2111 | ||
2112 | queue_for_each_hw_ctx(q, hctx, i) { | |
2113 | if (shared) | |
2114 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2115 | else | |
2116 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2117 | } | |
2118 | } | |
2119 | ||
2120 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2121 | { | |
2122 | struct request_queue *q; | |
2123 | ||
2124 | lockdep_assert_held(&set->tag_list_lock); | |
2125 | ||
2126 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2127 | blk_mq_freeze_queue(q); | |
2128 | queue_set_hctx_shared(q, shared); | |
2129 | blk_mq_unfreeze_queue(q); | |
2130 | } | |
2131 | } | |
2132 | ||
2133 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2134 | { | |
2135 | struct blk_mq_tag_set *set = q->tag_set; | |
2136 | ||
2137 | mutex_lock(&set->tag_list_lock); | |
2138 | list_del_rcu(&q->tag_set_list); | |
2139 | INIT_LIST_HEAD(&q->tag_set_list); | |
2140 | if (list_is_singular(&set->tag_list)) { | |
2141 | /* just transitioned to unshared */ | |
2142 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2143 | /* update existing queue */ | |
2144 | blk_mq_update_tag_set_depth(set, false); | |
2145 | } | |
2146 | mutex_unlock(&set->tag_list_lock); | |
2147 | ||
2148 | synchronize_rcu(); | |
2149 | } | |
2150 | ||
2151 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2152 | struct request_queue *q) | |
2153 | { | |
2154 | q->tag_set = set; | |
2155 | ||
2156 | mutex_lock(&set->tag_list_lock); | |
2157 | ||
2158 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2159 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2160 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2161 | /* update existing queue */ | |
2162 | blk_mq_update_tag_set_depth(set, true); | |
2163 | } | |
2164 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2165 | queue_set_hctx_shared(q, true); | |
2166 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); | |
2167 | ||
2168 | mutex_unlock(&set->tag_list_lock); | |
2169 | } | |
2170 | ||
2171 | /* | |
2172 | * It is the actual release handler for mq, but we do it from | |
2173 | * request queue's release handler for avoiding use-after-free | |
2174 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2175 | * but we can't group ctx/kctx kobj without it. | |
2176 | */ | |
2177 | void blk_mq_release(struct request_queue *q) | |
2178 | { | |
2179 | struct blk_mq_hw_ctx *hctx; | |
2180 | unsigned int i; | |
2181 | ||
2182 | /* hctx kobj stays in hctx */ | |
2183 | queue_for_each_hw_ctx(q, hctx, i) { | |
2184 | if (!hctx) | |
2185 | continue; | |
2186 | kobject_put(&hctx->kobj); | |
2187 | } | |
2188 | ||
2189 | q->mq_map = NULL; | |
2190 | ||
2191 | kfree(q->queue_hw_ctx); | |
2192 | ||
2193 | /* | |
2194 | * release .mq_kobj and sw queue's kobject now because | |
2195 | * both share lifetime with request queue. | |
2196 | */ | |
2197 | blk_mq_sysfs_deinit(q); | |
2198 | ||
2199 | free_percpu(q->queue_ctx); | |
2200 | } | |
2201 | ||
2202 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
2203 | { | |
2204 | struct request_queue *uninit_q, *q; | |
2205 | ||
2206 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2207 | if (!uninit_q) | |
2208 | return ERR_PTR(-ENOMEM); | |
2209 | ||
2210 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2211 | if (IS_ERR(q)) | |
2212 | blk_cleanup_queue(uninit_q); | |
2213 | ||
2214 | return q; | |
2215 | } | |
2216 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2217 | ||
2218 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, | |
2219 | struct request_queue *q) | |
2220 | { | |
2221 | int i, j; | |
2222 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
2223 | ||
2224 | blk_mq_sysfs_unregister(q); | |
2225 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2226 | int node; | |
2227 | ||
2228 | if (hctxs[i]) | |
2229 | continue; | |
2230 | ||
2231 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
2232 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), | |
2233 | GFP_KERNEL, node); | |
2234 | if (!hctxs[i]) | |
2235 | break; | |
2236 | ||
2237 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, | |
2238 | node)) { | |
2239 | kfree(hctxs[i]); | |
2240 | hctxs[i] = NULL; | |
2241 | break; | |
2242 | } | |
2243 | ||
2244 | atomic_set(&hctxs[i]->nr_active, 0); | |
2245 | hctxs[i]->numa_node = node; | |
2246 | hctxs[i]->queue_num = i; | |
2247 | ||
2248 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2249 | free_cpumask_var(hctxs[i]->cpumask); | |
2250 | kfree(hctxs[i]); | |
2251 | hctxs[i] = NULL; | |
2252 | break; | |
2253 | } | |
2254 | blk_mq_hctx_kobj_init(hctxs[i]); | |
2255 | } | |
2256 | for (j = i; j < q->nr_hw_queues; j++) { | |
2257 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2258 | ||
2259 | if (hctx) { | |
2260 | if (hctx->tags) | |
2261 | blk_mq_free_map_and_requests(set, j); | |
2262 | blk_mq_exit_hctx(q, set, hctx, j); | |
2263 | kobject_put(&hctx->kobj); | |
2264 | hctxs[j] = NULL; | |
2265 | ||
2266 | } | |
2267 | } | |
2268 | q->nr_hw_queues = i; | |
2269 | blk_mq_sysfs_register(q); | |
2270 | } | |
2271 | ||
2272 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2273 | struct request_queue *q) | |
2274 | { | |
2275 | /* mark the queue as mq asap */ | |
2276 | q->mq_ops = set->ops; | |
2277 | ||
2278 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, | |
2279 | blk_stat_rq_ddir, 2, q); | |
2280 | if (!q->poll_cb) | |
2281 | goto err_exit; | |
2282 | ||
2283 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2284 | if (!q->queue_ctx) | |
2285 | goto err_exit; | |
2286 | ||
2287 | /* init q->mq_kobj and sw queues' kobjects */ | |
2288 | blk_mq_sysfs_init(q); | |
2289 | ||
2290 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2291 | GFP_KERNEL, set->numa_node); | |
2292 | if (!q->queue_hw_ctx) | |
2293 | goto err_percpu; | |
2294 | ||
2295 | q->mq_map = set->mq_map; | |
2296 | ||
2297 | blk_mq_realloc_hw_ctxs(set, q); | |
2298 | if (!q->nr_hw_queues) | |
2299 | goto err_hctxs; | |
2300 | ||
2301 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); | |
2302 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); | |
2303 | ||
2304 | q->nr_queues = nr_cpu_ids; | |
2305 | ||
2306 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; | |
2307 | ||
2308 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) | |
2309 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2310 | ||
2311 | q->sg_reserved_size = INT_MAX; | |
2312 | ||
2313 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); | |
2314 | INIT_LIST_HEAD(&q->requeue_list); | |
2315 | spin_lock_init(&q->requeue_lock); | |
2316 | ||
2317 | blk_queue_make_request(q, blk_mq_make_request); | |
2318 | ||
2319 | /* | |
2320 | * Do this after blk_queue_make_request() overrides it... | |
2321 | */ | |
2322 | q->nr_requests = set->queue_depth; | |
2323 | ||
2324 | /* | |
2325 | * Default to classic polling | |
2326 | */ | |
2327 | q->poll_nsec = -1; | |
2328 | ||
2329 | if (set->ops->complete) | |
2330 | blk_queue_softirq_done(q, set->ops->complete); | |
2331 | ||
2332 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); | |
2333 | ||
2334 | get_online_cpus(); | |
2335 | mutex_lock(&all_q_mutex); | |
2336 | ||
2337 | list_add_tail(&q->all_q_node, &all_q_list); | |
2338 | blk_mq_add_queue_tag_set(set, q); | |
2339 | blk_mq_map_swqueue(q, cpu_online_mask); | |
2340 | ||
2341 | mutex_unlock(&all_q_mutex); | |
2342 | put_online_cpus(); | |
2343 | ||
2344 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { | |
2345 | int ret; | |
2346 | ||
2347 | ret = blk_mq_sched_init(q); | |
2348 | if (ret) | |
2349 | return ERR_PTR(ret); | |
2350 | } | |
2351 | ||
2352 | return q; | |
2353 | ||
2354 | err_hctxs: | |
2355 | kfree(q->queue_hw_ctx); | |
2356 | err_percpu: | |
2357 | free_percpu(q->queue_ctx); | |
2358 | err_exit: | |
2359 | q->mq_ops = NULL; | |
2360 | return ERR_PTR(-ENOMEM); | |
2361 | } | |
2362 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); | |
2363 | ||
2364 | void blk_mq_free_queue(struct request_queue *q) | |
2365 | { | |
2366 | struct blk_mq_tag_set *set = q->tag_set; | |
2367 | ||
2368 | mutex_lock(&all_q_mutex); | |
2369 | list_del_init(&q->all_q_node); | |
2370 | mutex_unlock(&all_q_mutex); | |
2371 | ||
2372 | blk_mq_del_queue_tag_set(q); | |
2373 | ||
2374 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); | |
2375 | } | |
2376 | ||
2377 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
2378 | static void blk_mq_queue_reinit(struct request_queue *q, | |
2379 | const struct cpumask *online_mask) | |
2380 | { | |
2381 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); | |
2382 | ||
2383 | blk_mq_sysfs_unregister(q); | |
2384 | ||
2385 | /* | |
2386 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2387 | * we should change hctx numa_node according to new topology (this | |
2388 | * involves free and re-allocate memory, worthy doing?) | |
2389 | */ | |
2390 | ||
2391 | blk_mq_map_swqueue(q, online_mask); | |
2392 | ||
2393 | blk_mq_sysfs_register(q); | |
2394 | } | |
2395 | ||
2396 | /* | |
2397 | * New online cpumask which is going to be set in this hotplug event. | |
2398 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2399 | * one-by-one and dynamically allocating this could result in a failure. | |
2400 | */ | |
2401 | static struct cpumask cpuhp_online_new; | |
2402 | ||
2403 | static void blk_mq_queue_reinit_work(void) | |
2404 | { | |
2405 | struct request_queue *q; | |
2406 | ||
2407 | mutex_lock(&all_q_mutex); | |
2408 | /* | |
2409 | * We need to freeze and reinit all existing queues. Freezing | |
2410 | * involves synchronous wait for an RCU grace period and doing it | |
2411 | * one by one may take a long time. Start freezing all queues in | |
2412 | * one swoop and then wait for the completions so that freezing can | |
2413 | * take place in parallel. | |
2414 | */ | |
2415 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2416 | blk_freeze_queue_start(q); | |
2417 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2418 | blk_mq_freeze_queue_wait(q); | |
2419 | ||
2420 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2421 | blk_mq_queue_reinit(q, &cpuhp_online_new); | |
2422 | ||
2423 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2424 | blk_mq_unfreeze_queue(q); | |
2425 | ||
2426 | mutex_unlock(&all_q_mutex); | |
2427 | } | |
2428 | ||
2429 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2430 | { | |
2431 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2432 | blk_mq_queue_reinit_work(); | |
2433 | return 0; | |
2434 | } | |
2435 | ||
2436 | /* | |
2437 | * Before hotadded cpu starts handling requests, new mappings must be | |
2438 | * established. Otherwise, these requests in hw queue might never be | |
2439 | * dispatched. | |
2440 | * | |
2441 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2442 | * for CPU0, and ctx1 for CPU1). | |
2443 | * | |
2444 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2445 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2446 | * | |
2447 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set | |
2448 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2449 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2450 | * ignored. | |
2451 | */ | |
2452 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2453 | { | |
2454 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2455 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2456 | blk_mq_queue_reinit_work(); | |
2457 | return 0; | |
2458 | } | |
2459 | ||
2460 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2461 | { | |
2462 | int i; | |
2463 | ||
2464 | for (i = 0; i < set->nr_hw_queues; i++) | |
2465 | if (!__blk_mq_alloc_rq_map(set, i)) | |
2466 | goto out_unwind; | |
2467 | ||
2468 | return 0; | |
2469 | ||
2470 | out_unwind: | |
2471 | while (--i >= 0) | |
2472 | blk_mq_free_rq_map(set->tags[i]); | |
2473 | ||
2474 | return -ENOMEM; | |
2475 | } | |
2476 | ||
2477 | /* | |
2478 | * Allocate the request maps associated with this tag_set. Note that this | |
2479 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2480 | * will be updated to reflect the allocated depth. | |
2481 | */ | |
2482 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2483 | { | |
2484 | unsigned int depth; | |
2485 | int err; | |
2486 | ||
2487 | depth = set->queue_depth; | |
2488 | do { | |
2489 | err = __blk_mq_alloc_rq_maps(set); | |
2490 | if (!err) | |
2491 | break; | |
2492 | ||
2493 | set->queue_depth >>= 1; | |
2494 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2495 | err = -ENOMEM; | |
2496 | break; | |
2497 | } | |
2498 | } while (set->queue_depth); | |
2499 | ||
2500 | if (!set->queue_depth || err) { | |
2501 | pr_err("blk-mq: failed to allocate request map\n"); | |
2502 | return -ENOMEM; | |
2503 | } | |
2504 | ||
2505 | if (depth != set->queue_depth) | |
2506 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2507 | depth, set->queue_depth); | |
2508 | ||
2509 | return 0; | |
2510 | } | |
2511 | ||
2512 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) | |
2513 | { | |
2514 | if (set->ops->map_queues) | |
2515 | return set->ops->map_queues(set); | |
2516 | else | |
2517 | return blk_mq_map_queues(set); | |
2518 | } | |
2519 | ||
2520 | /* | |
2521 | * Alloc a tag set to be associated with one or more request queues. | |
2522 | * May fail with EINVAL for various error conditions. May adjust the | |
2523 | * requested depth down, if if it too large. In that case, the set | |
2524 | * value will be stored in set->queue_depth. | |
2525 | */ | |
2526 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) | |
2527 | { | |
2528 | int ret; | |
2529 | ||
2530 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); | |
2531 | ||
2532 | if (!set->nr_hw_queues) | |
2533 | return -EINVAL; | |
2534 | if (!set->queue_depth) | |
2535 | return -EINVAL; | |
2536 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2537 | return -EINVAL; | |
2538 | ||
2539 | if (!set->ops->queue_rq) | |
2540 | return -EINVAL; | |
2541 | ||
2542 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { | |
2543 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2544 | BLK_MQ_MAX_DEPTH); | |
2545 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2546 | } | |
2547 | ||
2548 | /* | |
2549 | * If a crashdump is active, then we are potentially in a very | |
2550 | * memory constrained environment. Limit us to 1 queue and | |
2551 | * 64 tags to prevent using too much memory. | |
2552 | */ | |
2553 | if (is_kdump_kernel()) { | |
2554 | set->nr_hw_queues = 1; | |
2555 | set->queue_depth = min(64U, set->queue_depth); | |
2556 | } | |
2557 | /* | |
2558 | * There is no use for more h/w queues than cpus. | |
2559 | */ | |
2560 | if (set->nr_hw_queues > nr_cpu_ids) | |
2561 | set->nr_hw_queues = nr_cpu_ids; | |
2562 | ||
2563 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), | |
2564 | GFP_KERNEL, set->numa_node); | |
2565 | if (!set->tags) | |
2566 | return -ENOMEM; | |
2567 | ||
2568 | ret = -ENOMEM; | |
2569 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2570 | GFP_KERNEL, set->numa_node); | |
2571 | if (!set->mq_map) | |
2572 | goto out_free_tags; | |
2573 | ||
2574 | ret = blk_mq_update_queue_map(set); | |
2575 | if (ret) | |
2576 | goto out_free_mq_map; | |
2577 | ||
2578 | ret = blk_mq_alloc_rq_maps(set); | |
2579 | if (ret) | |
2580 | goto out_free_mq_map; | |
2581 | ||
2582 | mutex_init(&set->tag_list_lock); | |
2583 | INIT_LIST_HEAD(&set->tag_list); | |
2584 | ||
2585 | return 0; | |
2586 | ||
2587 | out_free_mq_map: | |
2588 | kfree(set->mq_map); | |
2589 | set->mq_map = NULL; | |
2590 | out_free_tags: | |
2591 | kfree(set->tags); | |
2592 | set->tags = NULL; | |
2593 | return ret; | |
2594 | } | |
2595 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2596 | ||
2597 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2598 | { | |
2599 | int i; | |
2600 | ||
2601 | for (i = 0; i < nr_cpu_ids; i++) | |
2602 | blk_mq_free_map_and_requests(set, i); | |
2603 | ||
2604 | kfree(set->mq_map); | |
2605 | set->mq_map = NULL; | |
2606 | ||
2607 | kfree(set->tags); | |
2608 | set->tags = NULL; | |
2609 | } | |
2610 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2611 | ||
2612 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) | |
2613 | { | |
2614 | struct blk_mq_tag_set *set = q->tag_set; | |
2615 | struct blk_mq_hw_ctx *hctx; | |
2616 | int i, ret; | |
2617 | ||
2618 | if (!set) | |
2619 | return -EINVAL; | |
2620 | ||
2621 | blk_mq_freeze_queue(q); | |
2622 | blk_mq_quiesce_queue(q); | |
2623 | ||
2624 | ret = 0; | |
2625 | queue_for_each_hw_ctx(q, hctx, i) { | |
2626 | if (!hctx->tags) | |
2627 | continue; | |
2628 | /* | |
2629 | * If we're using an MQ scheduler, just update the scheduler | |
2630 | * queue depth. This is similar to what the old code would do. | |
2631 | */ | |
2632 | if (!hctx->sched_tags) { | |
2633 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2634 | min(nr, set->queue_depth), | |
2635 | false); | |
2636 | } else { | |
2637 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2638 | nr, true); | |
2639 | } | |
2640 | if (ret) | |
2641 | break; | |
2642 | } | |
2643 | ||
2644 | if (!ret) | |
2645 | q->nr_requests = nr; | |
2646 | ||
2647 | blk_mq_unfreeze_queue(q); | |
2648 | blk_mq_start_stopped_hw_queues(q, true); | |
2649 | ||
2650 | return ret; | |
2651 | } | |
2652 | ||
2653 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2654 | { | |
2655 | struct request_queue *q; | |
2656 | ||
2657 | lockdep_assert_held(&set->tag_list_lock); | |
2658 | ||
2659 | if (nr_hw_queues > nr_cpu_ids) | |
2660 | nr_hw_queues = nr_cpu_ids; | |
2661 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2662 | return; | |
2663 | ||
2664 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2665 | blk_mq_freeze_queue(q); | |
2666 | ||
2667 | set->nr_hw_queues = nr_hw_queues; | |
2668 | blk_mq_update_queue_map(set); | |
2669 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2670 | blk_mq_realloc_hw_ctxs(set, q); | |
2671 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2672 | } | |
2673 | ||
2674 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2675 | blk_mq_unfreeze_queue(q); | |
2676 | } | |
2677 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2678 | ||
2679 | /* Enable polling stats and return whether they were already enabled. */ | |
2680 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2681 | { | |
2682 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2683 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2684 | return true; | |
2685 | blk_stat_add_callback(q, q->poll_cb); | |
2686 | return false; | |
2687 | } | |
2688 | ||
2689 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2690 | { | |
2691 | /* | |
2692 | * We don't arm the callback if polling stats are not enabled or the | |
2693 | * callback is already active. | |
2694 | */ | |
2695 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2696 | blk_stat_is_active(q->poll_cb)) | |
2697 | return; | |
2698 | ||
2699 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2700 | } | |
2701 | ||
2702 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2703 | { | |
2704 | struct request_queue *q = cb->data; | |
2705 | ||
2706 | if (cb->stat[READ].nr_samples) | |
2707 | q->poll_stat[READ] = cb->stat[READ]; | |
2708 | if (cb->stat[WRITE].nr_samples) | |
2709 | q->poll_stat[WRITE] = cb->stat[WRITE]; | |
2710 | } | |
2711 | ||
2712 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, | |
2713 | struct blk_mq_hw_ctx *hctx, | |
2714 | struct request *rq) | |
2715 | { | |
2716 | unsigned long ret = 0; | |
2717 | ||
2718 | /* | |
2719 | * If stats collection isn't on, don't sleep but turn it on for | |
2720 | * future users | |
2721 | */ | |
2722 | if (!blk_poll_stats_enable(q)) | |
2723 | return 0; | |
2724 | ||
2725 | /* | |
2726 | * As an optimistic guess, use half of the mean service time | |
2727 | * for this type of request. We can (and should) make this smarter. | |
2728 | * For instance, if the completion latencies are tight, we can | |
2729 | * get closer than just half the mean. This is especially | |
2730 | * important on devices where the completion latencies are longer | |
2731 | * than ~10 usec. | |
2732 | */ | |
2733 | if (req_op(rq) == REQ_OP_READ && q->poll_stat[READ].nr_samples) | |
2734 | ret = (q->poll_stat[READ].mean + 1) / 2; | |
2735 | else if (req_op(rq) == REQ_OP_WRITE && q->poll_stat[WRITE].nr_samples) | |
2736 | ret = (q->poll_stat[WRITE].mean + 1) / 2; | |
2737 | ||
2738 | return ret; | |
2739 | } | |
2740 | ||
2741 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, | |
2742 | struct blk_mq_hw_ctx *hctx, | |
2743 | struct request *rq) | |
2744 | { | |
2745 | struct hrtimer_sleeper hs; | |
2746 | enum hrtimer_mode mode; | |
2747 | unsigned int nsecs; | |
2748 | ktime_t kt; | |
2749 | ||
2750 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) | |
2751 | return false; | |
2752 | ||
2753 | /* | |
2754 | * poll_nsec can be: | |
2755 | * | |
2756 | * -1: don't ever hybrid sleep | |
2757 | * 0: use half of prev avg | |
2758 | * >0: use this specific value | |
2759 | */ | |
2760 | if (q->poll_nsec == -1) | |
2761 | return false; | |
2762 | else if (q->poll_nsec > 0) | |
2763 | nsecs = q->poll_nsec; | |
2764 | else | |
2765 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2766 | ||
2767 | if (!nsecs) | |
2768 | return false; | |
2769 | ||
2770 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2771 | ||
2772 | /* | |
2773 | * This will be replaced with the stats tracking code, using | |
2774 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2775 | */ | |
2776 | kt = nsecs; | |
2777 | ||
2778 | mode = HRTIMER_MODE_REL; | |
2779 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2780 | hrtimer_set_expires(&hs.timer, kt); | |
2781 | ||
2782 | hrtimer_init_sleeper(&hs, current); | |
2783 | do { | |
2784 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2785 | break; | |
2786 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2787 | hrtimer_start_expires(&hs.timer, mode); | |
2788 | if (hs.task) | |
2789 | io_schedule(); | |
2790 | hrtimer_cancel(&hs.timer); | |
2791 | mode = HRTIMER_MODE_ABS; | |
2792 | } while (hs.task && !signal_pending(current)); | |
2793 | ||
2794 | __set_current_state(TASK_RUNNING); | |
2795 | destroy_hrtimer_on_stack(&hs.timer); | |
2796 | return true; | |
2797 | } | |
2798 | ||
2799 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) | |
2800 | { | |
2801 | struct request_queue *q = hctx->queue; | |
2802 | long state; | |
2803 | ||
2804 | /* | |
2805 | * If we sleep, have the caller restart the poll loop to reset | |
2806 | * the state. Like for the other success return cases, the | |
2807 | * caller is responsible for checking if the IO completed. If | |
2808 | * the IO isn't complete, we'll get called again and will go | |
2809 | * straight to the busy poll loop. | |
2810 | */ | |
2811 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) | |
2812 | return true; | |
2813 | ||
2814 | hctx->poll_considered++; | |
2815 | ||
2816 | state = current->state; | |
2817 | while (!need_resched()) { | |
2818 | int ret; | |
2819 | ||
2820 | hctx->poll_invoked++; | |
2821 | ||
2822 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2823 | if (ret > 0) { | |
2824 | hctx->poll_success++; | |
2825 | set_current_state(TASK_RUNNING); | |
2826 | return true; | |
2827 | } | |
2828 | ||
2829 | if (signal_pending_state(state, current)) | |
2830 | set_current_state(TASK_RUNNING); | |
2831 | ||
2832 | if (current->state == TASK_RUNNING) | |
2833 | return true; | |
2834 | if (ret < 0) | |
2835 | break; | |
2836 | cpu_relax(); | |
2837 | } | |
2838 | ||
2839 | return false; | |
2840 | } | |
2841 | ||
2842 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2843 | { | |
2844 | struct blk_mq_hw_ctx *hctx; | |
2845 | struct blk_plug *plug; | |
2846 | struct request *rq; | |
2847 | ||
2848 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2849 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2850 | return false; | |
2851 | ||
2852 | plug = current->plug; | |
2853 | if (plug) | |
2854 | blk_flush_plug_list(plug, false); | |
2855 | ||
2856 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
2857 | if (!blk_qc_t_is_internal(cookie)) | |
2858 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2859 | else | |
2860 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
2861 | ||
2862 | return __blk_mq_poll(hctx, rq); | |
2863 | } | |
2864 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2865 | ||
2866 | void blk_mq_disable_hotplug(void) | |
2867 | { | |
2868 | mutex_lock(&all_q_mutex); | |
2869 | } | |
2870 | ||
2871 | void blk_mq_enable_hotplug(void) | |
2872 | { | |
2873 | mutex_unlock(&all_q_mutex); | |
2874 | } | |
2875 | ||
2876 | static int __init blk_mq_init(void) | |
2877 | { | |
2878 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, | |
2879 | blk_mq_hctx_notify_dead); | |
2880 | ||
2881 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", | |
2882 | blk_mq_queue_reinit_prepare, | |
2883 | blk_mq_queue_reinit_dead); | |
2884 | return 0; | |
2885 | } | |
2886 | subsys_initcall(blk_mq_init); |